Can LLMs Help Localize Fake Words in Partially Fake Speech?

A team of researchers from institutions including Johns Hopkins University has published a paper investigating whether text-trained Large Language Models (LLMs) can be used to detect and locate fake words within partially manipulated audio. The core idea is to build a speech LLM that performs fake word localization via next token prediction, essentially training the model to identify inconsistencies that signal an edit. This approach targets a sophisticated threat: audio where only specific words have been swapped or altered, leaving the rest of the speech genuine.

Experiments on the AV-Deepfake1M and PartialEdit databases revealed that the model frequently learns to leverage specific editing-style patterns from its training data as cues. A key finding is its reliance on detecting 'word-level polarity substitutions'—for example, swapping 'good' for 'bad'—which are common editing tactics in these datasets. This allows the model to perform well in controlled, in-domain scenarios where such patterns are present.

However, the research highlights a significant open question: the model's potential over-reliance on these particular, learned patterns. While effective against known editing styles, this approach may struggle to generalize to novel, unseen manipulation techniques that don't follow the same substitution rules. The paper, submitted to Interspeech 2026, frames this generalization challenge as a critical next step for developing robust, real-world audio deepfake detectors that can't be easily fooled by new attack methods.